arjunvekariyagithub / camelyon16-grand-challenge Goto Github PK

def get_bbox(cont_img, rgb_image=None):
  | _, contours, _ = cv2.findContours(cont_img, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

cv2.findContours return 2 parameters,Not 3 parameters.

In extract_patches.py:
def find_roi_bbox_tumor_gt_mask(self,mask_image):
bounding_boxes=self.get_bbox(np.array(mask))
But
def get_bbox(cont_img,image),has two parameters....
Is it a bug???

In this class,many function have a line 'patch.close()',Actually patch don't have close() Attr

,contours,=cv2.findContours(cont_img,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)

error: (-210) [Start]FindContours supports only CV_8UC1 images when mode != CV_RETR_FLOODFILL otherwise supports CV_32SC1 images only in function cvStartFindContours

Thank your for the good work, Is there any possibility that you can share the trained model/checkpoint.

would save me a lot of time, thanks!

Hi, I noticed that the mask used in the work is different from the original dataset ('tif' instead of 'png'), can I know how did you generate the mask files as 'tif' format?

In class PatchExtractor(object):

for bounding_box in bounding_boxes:
b_x_start = int(bounding_box[0])
  b_y_start = int(bounding_box[1])
  b_x_end = int(bounding_box[0]) + int(bounding_box[2])
  b_y_end = int(bounding_box[1]) + int(bounding_box[3])
  X = np.random.random_integers(b_x_start, high=b_x_end,
size=utils.NUM_POSITIVE_PATCHES_FROM_EACH_BBOX)
  Y = np.random.random_integers
(b_y_start,high=b_y_end,size=utils.NUM_POSITIVE_PATCHES_FROM_EACH_BBOX)
  for x, y in zip(X, Y):
  if int(tumor_gt_mask[y, x]) is utils.PIXEL_WHITE:
  patch = wsi_image.read_region((x * mag_factor, y * mag_factor), 0, (utils.PATCH_SIZE, utils.PATCH_SIZE))

(x,y) in bounding_box,But read_region can't ensure all patches in bounding_box...
I think you should do following check:
if(x+utils.PATCH_SIZE<b_x_end&&y+utils.PATCH_SIZE<b_y_end&&int(tumor_gt_mask[y,x] is
utils.PIXEL_WHITE):
patch=wsi_iamge.read_region()

What was the training time with Nvidia K40?

Hi Arjun. Great work.
Although, I do not get that why you have included only normal tissue patches in your validation set and have not included a single tumor tissue patch? The inception_eval.py evaluates the precision for false positive cases only and does not evaluate the errors that might come due to the false negative cases. Is there a specific rationale behind this that I am missing. Thanks.

Many people use some algorithm(like OTSU) to remove backgroud of the image?
But why do you not remove backgroud?